Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Agent-based Modeling for Business Problem Solving

Lecture 0

Vipin P. Veetil†

† Assistant Professor, Indian Institute of Management Kozhikode

LinkedIn ABM Network

Group 9588199

www.vipinveetil.com

1 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Inside The General Electric Corporation

◮ The General Electric Corporation is a collection of the quasi-structured

interactions between many workers pursuing their own goals and behaviors

◮ Goals and behaviors of of individual workers include minimizing effort, maximizing

department budgets, currying favors to superiors, and exaggerating their
contributions

◮ The goals and behaviors of one worker may conflict with that of another

◮ The profits and sales of GE emerge out of these interactions, some in harmony
with each other, others in conflict

2 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Exiting a stadium after a cricket match

◮ Some individuals head out alone, others in small groups like a family, yet others in
large groups like a college class

◮ People jostle around trying to get out soon while staying with their small group

◮ Some synergy between the goals of different individuals: the person ahead of you
moving forward helps you exit

◮ Some conflict between goals: it is easier for you to exit if the person in front of
you disappeared!

◮ The dynamics of the crowd emerges from the interactions between people with
goals that are in harmony and in conflict

3 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Spread of COVID

◮ A lockdown is announced

◮ The lockdown reduces interaction between people: lower virus spread

◮ People crowd to purchase supplies just before the lockdown: higher virus spread

◮ The impact of the lockdown ultimately depends on how the increase in spread
just before the lockdown plays out during the lockdown

◮ Dynamics of COVID emerges from the interactions between self-oriented

individuals before and after the lockdown

4 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Would adding more lanes to a road increase flow?

◮ Each car is less likely to be stuck behind a slow moving car

◮ You can move to a different lane and overtake

◮ Each car moves faster: more flow of traffic

◮ Each car disrupts the movement of other cars while changing lane
◮ Cars may have to slow down to let you in

◮ This slows some cars: less flow of traffic

◮ The dynamics of traffic flow emerges from the interactions between many
autonomous cars trying to do the best for themselves given the new lane

5 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Where to locate a new retail store?

◮ Many stores exist in regions where people live and shop

◮ Open a new store in a highly visited area or a sparse far-off area

◮ In highly visited area:

◮ Access to lots of potential customers

◮ But competitors may respond by cutting prices

◮ Your profits are very sensitive to how competitors respond

◮ Typically, this will not be a one time decision. Competitors will respond to you
and to each other. And you will respond to their response

◮ Collective dynamics of store location will depend on the interactions between

many such stores

6 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Common features in all five examples

◮ Many autonomous agents interact locally

◮ Collective outcome depends on interactions between individual actors pursuing their own
goals, which are sometimes in conflict with each other

◮ Each actor makes decisions based on what she knows about some others (like the speed
of the car ahead of her and the position of the car in the lane on the right).

◮ Actors do not have global information

◮ Out-of-equilibrium dynamics
◮ We would like to know how collective dynamics unfolds over time not just end-states
(equilibrium)

◮ We want to know how COVID virus spread changes in the immediate aftermath of the
lockdown, not after the rate of spread has settled into a steady state

◮ We want to know how the flow of people out of stadium is impacted in the immediate
aftermath of an explosion, not after the rate has settled into a new a steady state

7 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

When equation-based models, regressions, spreadsheets fail

◮ Very difficult to write down equations to describe the out-of-equilibrium flow of

traffic, spread of COVID after announcing a lockdown, and retail store dynamics
where stores choose prices and location

◮ Spreadsheets and regression analysis is not of much use either, they are pretty
poor at working with interactions between large number of actors

◮ Agent-based models come to the rescue :-)

8 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

What are agent-based models?

◮ Synthetic organization, powergrid, transport network, society, or economy in silicio

◮ A population of artificial agents who are:

◮ Autonomous

◮ Intelligent

◮ The artificial agents interact inside synthetic structure

9 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Empirical calibration of agent-based models?

◮ Calibrating a synthetic organization, power-grid, transport network, society, or

economy
◮ Match the number of power generators and the connections between then in a
distributed power-grid

◮ Recreate the spatial structure of the road network in a city while incorporating the
differences in the quality between different roads

◮ Calibrating the artificial agents

◮ Set the probability of an agent turning from the wine-isle to the pasta-isle to that which
is found in the empirical data from supermarket video footage

◮ Set the probability of a firm changing its input-supplier to that which is observed in the
data

◮ Can calibrate to individual agents or to population attributes: example risk-aversion

◮ Calibration depends on data availability

10 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Running a calibrated agent-based model

◮ Once the model is written in a programming language and calibrated to real world
data, it can be run forward in time

◮ Each time step, we can extract information on individual agents and the system
as a whole: the position of each individual exiting a stadium and the number of
individuals stuck at the exit point

◮ The data generated by the model over time is collected and analyzed to generate
insights

11 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Parametric variations

◮ Not all parameters can be accurately estimated with real world data
We may not have information on the probability of turning from the wine-aisle to the
pasta-aisle

◮ Agents behavior can change in response to our intervention

Probability of turning from the wine-aisle to the pasta-aisle in a new supermarket may
be different from that in the old

◮ The ABM is run forward in time for different parameteric specifications and data
collected from these runs are used to test the robustness of our insights
Run the model forward in time for probabilities of turning from the wine-aisle to the
pasta-aisle equal to 0.1, 0.5, and 0.9

Many runs are needed whenever agent behavior is stochastic behavior

12 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Agent-based models are testbeds in silico

13 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Why make decisions in silico before you make them in the real world

◮ Think through how people will respond

Creating an ABM makes you think through how people will respond to your new
idea, and how each person would respond to the response of others. Ultimately,
the impact of your new idea on the bottom-line depends on how people respond
to it and to each other.

◮ Make cheap mistakes

A bad idea implement on an artificial organization costs a few days of computing

resources. A bad idea in the real world costs jobs, profits, even lives.

◮ Prepare for the future

Simulating an organization or a supply-chain forward in time allows you to see

various possibilities that can arise in the future because of your new decisions
today. Which in turn creates the opportunity to prepare for the future.

14 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Did I try it in silico first?

15 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

'Getting on' with ABM

◮ A general purpose programing language

◮ Model thinking

◮ Zen

16 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

A general purpose programing language

◮ Python, Java, C++ (not Stata or Matlab)

Numerous ABM toolkits based on Python, Java, and C++ like Mesa, MASON, Repast
Toolkits help with visualizing data and organizing code
Toolkits constraint thinking, for good and for bad
Avoid these for now
Good to learn how to build ABM from scratch

◮ Python
Easier to learn
Code reads like text

◮ C++
10x faster
10x less memory

17 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Model thinking

◮ The model in your head:

Parts of a system
The interaction between the parts

◮ The model on a whiteboard:

Details about the parts
Specific rules about their interactions

◮ Barebones model in code:

Agents file
Structure file
Parameters file

18 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Zen

"...
perfection is finally attained not when there is no longer anything to
add, but when there is no longer anything to take away, when a body has
been stripped down to its nakedness"

De Saint-Exupéry speaking of aeroplanes

19 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Lectures

◮ Lectures

◮ Tutorial for support

20 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Lectures

◮ First half: talk on ABM

◮ Second half: build an ABM

21 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Tutorial

◮ Conceptual ABM tutorials

Resolving queries on how to build and implement ABM

◮ Coding ABM tutorials

Git

Profiling

Debugging

Parallel computing

22 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Big Data

◮ Big-data can be used to calibrate agent-based models

◮ Agent-based models can generate big-data

23 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Artificial Intelligence

◮ The agents that populate an ABM can be 'intelligent'

◮ Early generation models worked with zero-intelligence agents

◮ Some models are populated with agents that follow heuristics

◮ Agents can make decisions using neural networks and genetic algorithms

◮ Individual agents can co-evolve with the system

Open the hood

◮ Computational challenge with millions of interacting 'deep' agents

24 / 25
Emergence ABM In Silico First Learning ABM Course Plan ABM-BigData-AI Future

Agent-based Models as testbeds for business decision-making

◮ Companies and government running extensive calibrated models in which

decisions are tested before they are implement in the real world

◮ High demand for individuals who can build, use, and work with such models

◮ Potentially better world because of better decisions, terribly bad decisions avoided
because their consequences are seen in silica

◮ Beyond AI: what happens when many intelligent beings interact without a central
controller?

◮ Be the revolution!

25 / 25